8 Extreme points and the Krein–Milman theorem The next four chapters will focus on an important geometric aspect of compact sets, namely, the role of extreme points where: Definition An extreme point of a convex set, A, is a point x ∈ A, with the property that if x = θy +(1− θ)z with y,z ∈ A and θ ∈ [0, 1], then y = x and/or z = x. E(A) will denote the set of extreme points of A. In other words, an extreme point is a point that is not an interior point of any line segment lying entirely in A. This chapter will prove a point is a limit of convex combinations of extreme points and the following chapters will refine this repre- sentation of a general point. ν +1 Example 8.1 The ν-simplex, ∆ν , is given by (5.3) as the convex hull in R of {δ1 ,...,δν +1}, the coordinate vectors. It is easy to see its extreme points are { }ν +1 { ∈ Rν || |≤ } precisely the ν +1points δj j=1 . The hypercube C0 = x xi 1 has the 2ν points (±1, ±1,...,±1) as extreme points. The ball Bν = {x/∈ Rν ||x|≤ 1} has the entire sphere as extreme points, showing E(A) can be infinite. An interesting example (see Figure 8.1) is the set A ⊂ R3 , which is the convex hull of A = ch({(x, y, 0) | x2 + y2 =1}∪{(1, 0, ±1)}) (8.1) Its extreme points are E(A)={(x, y, 0) | x2 + y2 =1,x=1 }∪{(1, 0, ±1)} 1 1 − (1, 0, 0) = 2 (1, 0, 1) + 2 (1, 0, 1) is not an extreme point. This example shows that even in the finite-dimensional case, the extreme points may not be closed. In the infinite-dimensional case, we will even see that the set of extreme points can be dense! Extreme points and the Krein–Milman theorem 121 Not an extreme point Figure 8.1 An example of not closed extreme points If a point, x,inA is not extreme, it is an interior point of some segment [y,z]={θy +(1− θ)z | 0 ≤ θ ≤ 1} (8.2) with y = z.Ify or z is not an extreme point, we can write them as convex com- binations and continue. (If A is compact and in Rν , and if one extends the line segment to be maximal, one can prove this process will stop in finitely many steps. Indeed, that in essence is the method of proof we will use in Theorem 8.11.) If one thinks about writing y,z as convex combinations, one “expects” that any point in A is a convex linear combination of extreme points of A – and we will prove this when A is compact and finite-dimensional. Indeed, if A ⊂ Rν , we will prove that at most ν +1extreme points are needed. This fails in infinite dimension, but we will find a replacement, the Krein–Milman theorem, which says that any point is a limit of convex combinations of extreme points. These are the two main results of this chapter. Extreme points are a special case of a more general notion: Definition A face of a convex set is a nonempty subset, F ,ofA with the property that if x, y ∈ A, θ ∈ (0, 1), and θx +(1− θ)y ∈ F , then x, y ∈ F . A face, F , that is strictly smaller than A is called a proper face. Thus, a face is a subset so that any line segment [xz] ⊂ A, with interior points in F must lie in F . Extreme points are precisely one-point faces of A.(Note: See the remark before Proposition 8.6 for a later restriction of this definition.) Example 8.2 (Example 8.1 continued) ∆ν has lots of faces; explicitly, it has ν +1 − ν +1 ν +1 2 2 proper faces, namely, ν +1extreme points, 2 facial lines, ..., ν − ν − ν faces of dimension (ν 1). The hypercube Cν has 3 1 faces, namely, 2 ex- ν −1 ν ν −2 ν treme points, ν2 facial lines, 2 2 facial planes, ..., 2 ν −1 faces of di- mension (ν − 1). The only faces on the ball are its extreme points. The faces of the set A of (8.1) are its extreme points, the line {(1, 0,y) ||y|≤1}, and the lines 122 Convexity {θ(x0 ,y0 , 0)+(1−θ)(1, 0, 1)} and {θ(x0 ,y0 , 0)+(1−θ)(1, 0, −1)}, where x0 ,y0 2 2 are fixed with x0 + y0 =1and x0 =1. A canonical way proper faces are constructed is via linear functionals. Theorem 8.3 Let A be a convex subset of a real vector space. Let : A → R be a linear functional with (i) sup (x)=α<∞ (8.3) x∈A (ii) A is not constant. Then {y | (y)=α} = F (8.4) if nonempty, is a proper face of A. Remark If A is compact and is continuous, of course, F is nonempty. Proof Since is linear, F is convex. Moreover, if y,z ∈ A and θ ∈ (0, 1) and θy +(1− θ)z ∈ F , then θ(y)+(1− θ)(z)=α and (y) ≤ α, (z) ≤ α implies (y)=(z)=α, that is, y,z ∈ F . By (ii), F is a proper subset of A. The hyperplane {y | (y)=α} with α given by (8.3) is called a tangent hyper- plane or support hyperplane. The set (8.4) is called an exposed set.IfF is a single point, we call the point an exposed point. Example 8.4 We have just seen that every exposed set is a face so, in particular, every exposed point is an extreme point. I’ll bet if you think through a few simple examples like a disk or triangle in the plane or a convex polyhedron in R3 , you’ll conjecture the converse is true. But it is not! Here is a counterexample in R2 (see Figure 8.2): A = {(x, y) |−1 ≤ x ≤ 1, −2 ≤ y ≤ 0}∪{(x, y) | x2 + y2 ≤ 1} The boundary of A above y = −2 is a C1 curve, so there is a unique supporting hyperplane through each such boundary point. The supporting hyperplane through the extreme point (1, 0) is x =1so (1, 0) is not an exposed point, but it is an extreme point. Proposition 8.5 Any proper face F of A lies in the topological boundary of A. Conversely, if A ⊂ X, a locally convex space (and, in particular, in Rν ), and Aint is nonempty, then any point x ∈ A ∩ ∂A lies in a proper face. Extreme points and the Krein–Milman theorem 123 A nonexposed extreme point Figure 8.2 A nonexposed extreme point Proof Let x ∈ F and pick y ∈ A\F .Thesetofθ ∈ R so z(θ) ≡ θx+(1− θ)y ∈ A includes [0, 1], but it cannot include any θ>1 for if it did, θ =1(i.e., x) would be an interior point of a line in A with at least one endpoint in A\F . Thus, x = −1 limn↓0 z(1 + n ) is a limit point of points not in A, that is, x ∈ A¯ ∩ X\A = ∂A. For the converse, let x ∈ A∩∂A and let B = Aint. Since B is open, Theorem 4.1 ≤ implies there exists a continuous L =0with α =supy∈B L(y) L(x). Since x ∈ A, L(x)=α. Since B is open, L[B] is an open set (Lemma 4.2), so the supporting hyperplane H = {y | L(y)=α} is disjoint from B and so H ∩ A is a proper face. To have lots of extreme points, we will need lots of boundary points, so it is Rν { ∈ R | natural to restrict ourselves to closed convex sets. The convex set + = x xi ≥ 0 all i} has a single extreme point, so we will also restrict to bounded sets. Indeed, except for some examples, we will restrict ourselves to compact convex sets in the infinite-dimensional case. Convex cones are interesting but can normally be treated as suspensions of compact convex sets; see the discussion in Chapter 11. So we will suppose A is a compact convex subset of a locally convex space. As noted in Corollary 4.9, A is weakly compact, so we will suppose henceforth that we are dealing with the weak topology. Remark Henceforth, we will also restrict the term “face” to indicate a closed set. Proposition 8.6 Let F ⊂ A with A a compact convex set and F a face of A. Let B ⊂ F . Then B is a face of F if and only if it is a face of A. In particular, x ∈ F is in E(F ) if and only if it is also in E(A), that is, E(F )=F ∩ E(A) Proof If B is a face of A, x ∈ B, and x is an interior point of [y,z] ⊂ F ,itisan interior point of [y,z] ⊂ A. Thus, y,z ∈ A,soy,z ∈ B, and thus, B is a face of F . Conversely, if B is a face of F , x ∈ B, and x ∈ [y,z] ⊂ A, since x ∈ F ,the fact that F is a face implies y,z ∈ F so [y,z] ⊂ F . Thus, since B is a face of F , y,z ∈ B and so B is a face of A. 124 Convexity We turn next to a detailed study of the finite-dimensional case. We begin with some notions that involve finite dimension but which are useful in the infinite- dimensional case also.